SIMD [iu]8xN tests SIGILL in --release #88769
Labels
A-LLVM
Area: Code generation parts specific to LLVM. Both correctness bugs and optimization-related issues.
A-SIMD
Area: SIMD (Single Instruction Multiple Data)
C-bug
Category: This is a bug.
E-needs-test
Call for participation: An issue has been fixed and does not reproduce, but no test has been added.
I-crash
Issue: The compiler crashes (SIGSEGV, SIGABRT, etc). Use I-ICE instead when the compiler panics.
I-monomorphization
Issue: An error at monomorphization time.
PG-portable-simd
Project group: Portable SIMD (https://github.com/rust-lang/project-portable-simd)
requires-nightly
This issue requires a nightly compiler in some way.
Comments
|
During my reduction, this nice error popped up which is, at least I think so, the real error message behind this ICE: I post my code in ~10 Minutes, after cleaning it up |
|
This ICE is very fragile, even removing the #![feature(platform_intrinsics, repr_simd)]
extern "platform-intrinsic" {
pub(crate) fn simd_reduce_mul_ordered<T, U>(x: T, y: U) -> U;
}
#[repr(simd)]
pub struct Simd<T, const LANES: usize>([T; LANES]);
impl<T, const LANES: usize> Simd<T, LANES> {
pub fn new(array: [T; LANES]) -> Self {
Self(array)
}
}
mod biteq {
pub trait BitEq {
fn biteq(&self, other: &Self) -> bool;
}
impl BitEq for i8 {
fn biteq(&self, _: &Self) -> bool {
false
}
}
pub struct BitEqWrapper<T>(pub T);
impl<T: BitEq> PartialEq for BitEqWrapper<T> {
fn eq(&self, _: &Self) -> bool {
false
}
}
}
fn main() {
pub fn implementation(x: [i8; 16]) {
use crate::biteq::BitEqWrapper;
let a = unsafe { crate::simd_reduce_mul_ordered(crate::Simd::<i8, 16>::new(x), 1) };
let b = 0;
let left = BitEqWrapper(a);
let right = BitEqWrapper(b);
let _ = left == right;
}
implementation([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]);
}
BacktraceGdb: based on the error message: @rustbot modify labels: I-crash I-monomorphization |
is the assertion being triggered (with @hellow554's minimization) |
|
I do not understand how that is finding a null pointer or how it is getting a poison value out of this. |
|
Thank you for your work on the minimization though! I just... is it because it starts as an array and then it tries to constant fold that? And why only vectors of 8-bit values? The test is the same for larger ones. |
|
As a note, this crash only reproduces with |
|
I can repro with this: But only on |
|
Running with These series of steps reproduce the crash outside mkdir issue-88769
cd issue-88769
curl -sSO https://gist.githubusercontent.com/eddyb/08ea4736082d9b2326a984c4cb828a08/raw/6a2b96cd8422950fa2b160e518671df66965a863/issue-88769-min.rs
rustup toolchain install nightly-2021-09-08 -c llvm-tools-preview
# This will crash - ignore the crash and keep going, `-C save-temps` works.
rustc +nightly-2021-09-08 issue-88769-min.rs -O -C save-temps
# This reproduces the crash using the `.bc` from `-C save-temps`.
~/.rustup/toolchains/nightly-2021-09-08-x*64*/lib/rustlib/x*64*/bin/opt -O3 issue-88769-min.*-cgu.3.rcgu.no-opt.bc > opt.bcSadly we don't ship bugpoint --opt-command=$(echo ~/.rustup/toolchains/nightly-2021-09-08-x*64*/lib/rustlib/x*64*/bin/opt) -O3 issue-88769-min.*-cgu.3.rcgu.no-opt.bcAfter a lot of crashes, it ends in: Running ; ModuleID = 'bugpoint-reduced-simplified.bc'
source_filename = "issue_88769_min.61742fa9-cgu.3"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"
define dso_local void @_ZN15issue_88769_min4main17h9167823f857de443E() unnamed_addr #0 {
%1 = call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> bitcast (<1 x i128> <i128 20011376718272490338853433276725592320> to <16 x i8>))
store i8 %1, i8* undef, align 1
ret void
}
; Function Attrs: nofree nosync nounwind readnone willreturn
declare i8 @llvm.vector.reduce.mul.v16i8(<16 x i8>) #1
attributes #0 = { "target-cpu"="x86-64" }
attributes #1 = { nofree nosync nounwind readnone willreturn }
!llvm.module.flags = !{!0}
!0 = !{i32 2, !"RtLibUseGOT", i32 1}cc @nikic ( |
|
Just to confirm: crash repros on mkdir issue-88769
cd issue-88769
curl -sSO https://gist.githubusercontent.com/eddyb/08ea4736082d9b2326a984c4cb828a08/raw/6a2b96cd8422950fa2b160e518671df66965a863/issue-88769-min.rs
rustup toolchain install nightly-2021-08-22 -c llvm-tools-preview
# This will crash - ignore the crash and keep going, `-C save-temps` works.
rustc +nightly-2021-08-22 issue-88769-min.rs -O -C save-temps
# This reproduces the crash using the `.bc` from `-C save-temps`.
~/.rustup/toolchains/nightly-2021-08-22-x*64*/lib/rustlib/x*64*/bin/opt -O3 issue-88769-min.*-cgu.3.rcgu.no-opt.bc > opt.bc |
|
@eddyb Thanks for the reduction, upstream issue filed in https://bugs.llvm.org/show_bug.cgi?id=51811. |
Merged
|
Should be fixed by #88765. |
|
Okay, trying to compile our test still appears to crash |
|
The LLVM-level repro and hellow's minimization both seem to be fixed. |
|
The reduction above now works on nightly, but portable-simd still fails on the same test with the same error as the original report. I'm guessing the bug is only partially fixed in LLVM? |
|
I can confirm that my reduced testcase is indeed fixed, but portable-simd still fails on linux and windows. I'll try it again :) It seems not to be a llvm issue anymore, but a rust (jobserver) one? It panicks with a SIGTRAP under linux, here's the backtrace: thread1: thread2: |
|
MCVE: #![feature(platform_intrinsics, repr_simd)]
#[repr(simd)]
pub struct Simd<T, const LANES: usize>([T; LANES]);
impl<const LANES: usize> Simd<i8, LANES> {
pub fn horizontal_product(self) -> i8 {
extern "platform-intrinsic" {
fn simd_reduce_mul_ordered<T, U>(x: T, y: U) -> U;
}
unsafe { simd_reduce_mul_ordered(self, 1) }
}
}
fn cb(x: [i8; 16]) {
let a: i8 = Simd(x).horizontal_product();
let _ = a.to_string();
}
fn main() {
let x = [0; 16];
let _ = std::panic::catch_unwind(|| cb(x));
}Regressed in db002a0 so it is an llvm issue after all? |
|
Reduced: |
|
Upstream report: https://bugs.llvm.org/show_bug.cgi?id=51858 |
|
What appears to be a fix? is now present upstream at llvm/llvm-project@dcba994 |
Merged
bors
added a commit
to rust-lang-ci/rust
that referenced
this issue
Sep 24, 2021
Update LLVM submodule This merges the upstream `release/13.x` branch to pull in the second fix for rust-lang#88769.
|
Should be fixed by #89130 -- unless there's a third bug? |
|
I re-ran a test and it appears fixed, thanks! I'll wait for @workingjubilee to confirm. |
|
Ditto! No third bug this time. 😅 Thank you so much! |
Merged
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Hello, we have encountered an issue in the portable-simd CI unique to tests on our
i8xNandu8xNtypes on x86_64. It emits a SIGTRAP, or sometimes a SIGILL, or on Windows, a STATUS_ACCESS_VIOLATION! How exciting! I can reproduce this locally. Our other tests do not do this, and it only happens in--release, so I suspect this is a miscompilation by LLVM 13 inappropriately optimizing something. Full command to repro:cargo test --release --test [iu]8_opsI am going to try to minimize this from "literally the entire test suite" but I might take a bit. The issue became apparent about 7 days ago. Tests were passing about 20
30 days ago (August 818) judging by the last green in rust-lang/portable-simd#155. I believe the time frame is correct for this to have been due to #87570More minutiae
RUSTFLAGS="-Copt-level=1"passes with the above tests.RUSTFLAGS="-Copt-level=2"crashes with them.Version it worked on
It most recently worked on: Rust 1.56 nightly (2021-07-17)
Version with regression
rustc --version --verbose:but it also happened on rustc 1.57.0-nightly 2021-09-01
Sample Crash
The text was updated successfully, but these errors were encountered: